- GoHighLevel's official MCP server connects Claude to your CRM in about ten minutes, free on every Claude plan.
- The 2026 endpoint replaced the old 36-tool list with five tools that reach hundreds of operations across your account.
- Most failed setups are one of five errors: wrong key type, missing scopes, wrong Location ID, bad record IDs, or config typos.
- Claude reads and writes records but can't build workflows. That stays human work for now.
- Claude with MCP is an operator copilot; the $97/month AI Employee is customer-facing. Different jobs, both useful.
What does connecting Claude to GoHighLevel actually mean?
HighLevel runs an official MCP server: a live connection that lets Claude read and act on your CRM in plain English. Contacts, conversations, opportunities, calendars, invoices. You ask, it answers or acts. There's no Zapier in the middle and nothing to host. HighLevel launched the server on July 30, 2025 with 21 tools and has expanded it ever since.
MCP, the Model Context Protocol, is the connection standard Anthropic open-sourced in November 2024. Within a year it passed 97 million monthly SDK downloads and 10,000 active servers. Its governance now sits with a foundation backed by Anthropic, OpenAI, Google, and Microsoft. In plain terms: this is the plug the AI industry settled on, not a plugin fad.
That matters because HighLevel reports more than a million small businesses on its platform, and most use a fraction of what they pay for. A connected assistant is the cheapest fix we know for that gap. In Salesforce's research, 91% of small businesses using AI say it boosts their revenue. (And if the account itself is still half-configured, start with our complete GoHighLevel setup guide before wiring in AI.)
The 2026 update most guides miss
If the tutorial you're reading says the server has '36 tools,' it's describing last year's setup. HighLevel's current docs point Claude users to a newer v2 endpoint: services.leadconnectorhq.com/mcp/anthropic/v2. It works differently. Five tools search, describe, and execute hundreds of operations across roughly 40 areas of your account.
The difference is practical, not academic. On the old endpoint, Claude had a fixed menu. On v2, it looks up the right operation for whatever you asked, then runs it: payments, social posts, blog content, form submissions. More of your account is reachable, and new operations appear without you touching anything.
Sign-in got simpler too. The docs now recommend OAuth: click connect, log into HighLevel in a browser window, approve, done. The older method still works everywhere: pasting a Private Integration Token into a request header. Most guides published this month still only teach the token method.
How do you set up the GoHighLevel MCP server?
You need three things: a GoHighLevel sub-account, either OAuth approval or a Private Integration Token, and Claude on any plan, including Free. Custom connectors are available on every Claude tier, so the connection itself costs nothing.

- In your sub-account, go to Settings → Private Integrations → Create. Select scopes — start with View Contacts, View Conversations, and View Opportunities — then copy the token. It starts with pit-.
- Grab your Location ID from Settings → Business Profile. It's also in your sub-account URL.
- In claude.ai or Claude Desktop, open Settings → Connectors → Add custom connector and paste the v2 endpoint. Choose OAuth if offered, or add the token as an Authorization header plus a locationId header.
- In Claude Code, it's one command: claude mcp add with the endpoint and the same two headers.
- Test it: 'List the contacts created this week.' You should see real names.
Why won't Claude connect? The five errors that break the setup
Nearly every failed setup we've seen traces to one of five causes. Check them in order and you'll fix most problems without opening a support ticket.
- 401: you used a regular API key. Only Private Integration Tokens work — the ones starting with pit-.
- 403: the token is real but missing a scope. Edit the integration, add the scope, save, and retry.
- Data from the wrong client account: your locationId header is missing or wrong. One token, one location.
- 404: usually a bad record ID, sometimes an operation name the server doesn't recognize.
- 'Server not found' on local community servers: almost always a config-file typo or path problem. The official remote endpoint skips that entire class of failure.
What we run through it every day
There are quieter blockers than the five above, too. Our team hit one this month: creating a developer account on GoHighLevel's marketplace requires a US phone number for the verification code. Our build team is global, so someone stateside had to create the account first. No doc we've found mentions it.
Once connected, the connection earns its keep in unglamorous work. Morning pipeline pulls. Stale-lead audits. Drafting follow-ups. The prompts that get used aren't clever. 'Find contacts who filled the inquiry form this week but haven't been contacted.' 'List opportunities with no activity in 14 days so we can tag them cold.' We keep the full list of ten MCP use cases we run on client accounts as its own guide.
If you want a starting routine, this is the one we'd hand a new client for week one:
The AI isn't the impressive part. The impressive part is answering questions about your own business without logging into anything.
- Monday: 'Summarize what changed in my pipeline last week and flag anything stalled.'
- Every morning: 'List today's appointments and any that are unconfirmed.'
- Midweek: 'Find contacts created this month with no follow-up yet, and draft a short re-engagement text for each.'
- Month-end: 'Pull closed-won opportunities by source, so we can see which campaigns actually paid.'
Why speed-to-lead is the whole game
Here's why the speed matters. In May we sat with a dental practice spending about $1,000 a month on ads and generating three or four implant leads a day. Those are $18,000-to-$36,000 cases. Follow-up happened the next business day, and weekend leads waited until Monday. Around 13,000 past patients sat untouched in the database.
That pattern is everywhere. A 2025 study tested how 1,333 law firms handled a web lead. Only 25% responded within five minutes; 26% never responded at all. An assistant that can see the CRM is how that gap starts closing.
What Claude can't do in GoHighLevel yet
Claude can't build workflows in GoHighLevel. Workflow access is read-only, and AI-built workflows are one of the most-upvoted requests on HighLevel's ideas board, at 262 votes and counting. It also can't see across sub-accounts: one connection covers one location, so an agency with 20 clients manages 20 scoped tokens.
When we needed more than the official server offers, like calendar free-slot lookups and workflow enrollment, we built a small custom integration on GoHighLevel's API. That's the kind of AI-assisted system we build for clients, so take the recommendation with that in mind.
Claude with GoHighLevel or the AI Employee: which do you need?
They're different hires. Claude with MCP is a copilot for you, the operator. It answers, audits, and drafts when you ask. HighLevel's AI Employee is customer-facing automation: Voice AI answers your phone, Conversation AI answers your messages, and it runs whether you're watching or not. We broke down whether the AI Employee is worth the $97 separately. Most businesses that get value from one eventually run both.
The pricing is public. AI Employee Unlimited costs $97 a month per location, with fair-use terms. Pay-per-use is the alternative; voice minutes start around $0.045 plus a text-to-speech rate. Through August 31, 2026, HighLevel's 'Summer of AI' promotion includes 30-day trials of the customer-facing tools. That's the cheap way to find out whether your call volume justifies it.
Budget tuning time either way. A lawn-care client of ours ran live test calls against their new Voice AI phone agent, and it failed the first round. It collected the caller's number too late and offered a manager too early. The fixes were small and specific: ask for a number up front 'in case we get disconnected,' lower the response randomness, cap calls at six to seven minutes. After that, it booked.
And the July YouTube wave claiming Claude 'killed' GoHighLevel is half true. Claude really can build you a custom CRM in an afternoon. Then you own its hosting, backups, telephony, and carrier compliance forever. For a business with real call volume and client data, that isn't savings. It's a second job.
- Claude + MCP: works for you, the operator. Reads and writes with scopes you choose. Free beyond the plans you already have.
- AI Employee: works for your customers. Answers calls and messages on its own. $97 a month per location, or pay-per-use.
- ChatGPT connector: official but read-only, needs a paid ChatGPT plan with developer mode, and HighLevel notes the data is shared with OpenAI.
Treat the token like a master key
A full-scope Private Integration Token can read and edit everything you granted: contacts, conversations, invoices, payments. Scope it the way you'd scope an employee login. Least privilege, one token per purpose, revoked the day it's not needed.
- Keep a read-only 'reporting' token for daily questions and audits.
- Create a separate write-enabled token only if you actually want Claude editing records, and only with the scopes that job needs.
- Name tokens by location and purpose so you can revoke one without breaking the rest.
- Rotate on staff changes. It takes a minute in Settings → Private Integrations.
- HighLevel API docs — LeadConnector MCP Server
- HighLevel Support — How to Use the HighLevel MCP Server
- GoHighLevel — Introducing the MCP Server (Jul 2025)
- HighLevel Support — AI Product Pricing (updated Jun 2026)
- HighLevel Support — Summer of AI 2026
- Anthropic — Introducing the Model Context Protocol (Nov 2024)
- MCP Blog — MCP Joins the Agentic AI Foundation (Dec 2025)
- Claude Help Center — Custom Connectors Using Remote MCP
- Salesforce — SMBs With AI Adoption See Stronger Revenue Growth
- Hennessey Digital — 2025 Lead Form Response Time Study
- HighLevel Ideas Board — Claude → GHL Workflow Automations (Write)

Samar runs Webly Studio, the agency behind the paid ads, web builds, and AI systems featured on this blog. The team's work and results live at /work.




